TL;DR
Cursor just launched Composer 2, its own proprietary coding model. It beats Claude Opus 4.6 on multiple benchmarks, costs 86% less than their previous model, and signals a massive strategic shift: the $29.3 billion coding startup is building its own AI models instead of just wrapping Anthropic and OpenAI. GPT-5.4 still leads on Terminal-Bench 2.0, but Cursor is closing the gap fast.
The Numbers
| Metric | Value |
|---|---|
| Model | Composer 2 |
| CursorBench score | 61.3 (vs 44.2 for Composer 1.5) |
| Terminal-Bench 2.0 | 61.7 (beats Claude Opus 4.6’s 58.0) |
| SWE-bench Multilingual | 73.7 (vs 65.9 for Composer 1.5) |
| Standard pricing | $0.50/M input, $2.50/M output |
| Fast pricing | $1.50/M input, $7.50/M output |
| Context window | 200,000 tokens |
| Price drop vs Composer 1.5 | ~86% cheaper |
| Available | Now, in Cursor + Glass alpha |
What Happened
On March 19, 2026, Cursor (Anysphere Inc.) launched Composer 2 — a proprietary AI coding model trained from scratch on coding-only datasets. This isn’t a fine-tuned version of someone else’s model. It’s Cursor’s own foundation model, built specifically for their agentic coding environment.
The release came alongside two other announcements:
- Composer 2 Fast — A higher-speed variant that’s now the default for all users
- Glass — A new interface for Cursor, currently in early alpha
Covered by Bloomberg, VentureBeat, and SiliconANGLE, the launch represents a strategic pivot from being a platform that wraps third-party models to becoming a vertically integrated AI coding company.
Benchmark Breakdown
CursorBench (Internal)
Based on real tasks from Cursor’s own engineering team — averaging 352 lines across 8 files per challenge:
| Model | Score |
|---|---|
| GPT-5.4 (high) | ~66 |
| GPT-5.4 (medium) | ~63 |
| Composer 2 | 61.3 |
| GPT-5.4 (low) | ~59 |
| Claude Opus 4.6 | ~57 |
| Composer 1.5 | 44.2 |
| Composer 1 | 38.0 |
Terminal-Bench 2.0 (Third-Party)
Measures AI agents’ ability to perform tasks in terminal environments:
| Model | Score |
|---|---|
| GPT-5.4 | 75.1 |
| Composer 2 | 61.7 |
| Claude Opus 4.6 | 58.0 |
| Claude Opus 4.5 | 52.1 |
| Composer 1.5 | 47.9 |
| Composer 1 | 40.0 |
SWE-bench Multilingual (Third-Party)
Multi-language software engineering tasks:
| Model | Score |
|---|---|
| Composer 2 | 73.7 |
| Composer 1.5 | 65.9 |
| Composer 1 | 56.9 |
Key takeaway: Composer 2 outperforms Claude Opus 4.6 and sits between GPT-5.4’s low and medium configurations. Not the best model on every benchmark — but competitive at a fraction of the cost.
The Pricing Story Is Equally Important
Composer 2 isn’t just better — it’s dramatically cheaper:
| Model | Input (per M tokens) | Output (per M tokens) |
|---|---|---|
| Composer 1.5 | $3.50 | $17.50 |
| Composer 2 | $0.50 | $2.50 |
| Composer 2 Fast | $1.50 | $7.50 |
That’s an 86% price reduction from Composer 1.5 on standard, and 57% cheaper even on the fast variant.
Why so cheap? According to Bloomberg, the cost-efficiency stems from training solely on coding datasets. Frontier models like GPT-5.4 and Claude Opus 4.6 are trained on everything — text, code, math, reasoning, images. A code-only model needs less data and less compute to reach competitive performance on coding tasks.
How They Built It
Cursor revealed several key technical details:
Continued Pretraining
Composer 2 is their first model built with continued pretraining — taking a base model and further training it on coding-specific data. This gave a “far stronger base” for reinforcement learning.
Reinforcement Learning on Long-Horizon Tasks
The model was trained on extended coding tasks requiring hundreds of actions — not just one-shot code generation. This means it can:
- Read an entire repository
- Plan multi-file changes
- Execute commands
- Interpret errors
- Iterate toward a solution
Self-Summarization
For tasks exceeding the context window, Cursor developed self-summarization — compressing information without losing critical details. This lets the model maintain context across very long coding sessions.
Tool Integration
Composer 2 has native access to Cursor’s full tool stack:
- Semantic code search
- File and folder operations
- Shell commands
- Browser control
- Web access
- Image generation
Why This Matters: The Vertical Integration Play
This launch signals something bigger than a new model. Cursor is becoming vertically integrated — owning the model, the IDE, the agent infrastructure, and the user relationship.
Before Composer 2
Developer → Cursor IDE → Claude/GPT API → ResponseCursor was a smart wrapper. Anthropic or OpenAI could ship their own coding tools at any time (and they did — Claude Code and OpenAI Codex exist).
After Composer 2
Developer → Cursor IDE → Composer 2 → ResponseCursor controls the full stack. They can optimize the model for their specific workflows, reduce costs, and differentiate in ways competitors can’t replicate.
The Strategic Threat
The move is defensive as much as offensive. VentureBeat reported growing social media chatter from developers switching from Cursor to Claude Code — especially power users who prefer terminal-first, fully agentic workflows.
As Anthropic and OpenAI ship their own coding products, Cursor faces an existential question: why pay for a platform when the model makers build their own?
Composer 2 is the answer: Cursor wants to prove it’s not just a wrapper. It’s building its own intelligence.
Glass: The New Interface
Alongside Composer 2, Cursor quietly launched Glass — a new interface in early alpha. Details are scarce (the community forums are already reporting bugs with WSL and C/C++ support), but it signals Cursor is rethinking the IDE experience beyond VS Code.
The combination of a proprietary model + a proprietary interface could create a fully differentiated product that’s harder for competitors to replicate.
How Composer 2 Compares to Claude Code
This is the comparison everyone’s making:
| Feature | Cursor + Composer 2 | Claude Code |
|---|---|---|
| Interface | IDE (VS Code fork) + Glass | Terminal |
| Model | Proprietary Composer 2 | Claude Opus 4.5/4.6 |
| Context window | 200K tokens | 200K tokens |
| Multi-file editing | Subagent parallelism | Sequential |
| Terminal access | Via IDE | Native |
| Pricing | $20/mo Pro, $60/mo Pro+ | $20/mo Claude Pro |
| Best for | Visual coding, team use | Complex reasoning, solo devs |
| Benchmark edge | Cost efficiency | Raw reasoning quality |
The honest take: Claude Code still has the edge on raw reasoning quality (Opus 4.5/4.6 are exceptional). But Composer 2 offers competitive performance at dramatically lower per-token costs, with tighter IDE integration.
Reddit user feedback reflects this: “I think a good workflow is to use both. Claude is nuts in terms of usage and can’t be beat but Cursor is much faster and having different models is great.”
What Developers Should Know
If You’re Already on Cursor
Composer 2 is available now. It’s the default in Fast mode. Try it — the 86% price reduction means you can use it far more aggressively.
If You’re on Claude Code
Don’t switch yet. Claude’s reasoning quality still leads. But watch Cursor’s trajectory — they’re closing the gap fast.
If You’re Choosing Between Them
The real answer is both:
- Cursor + Composer 2 for day-to-day coding, UI work, quick iterations
- Claude Code for complex debugging, architecture decisions, large refactors
Pricing Quick Reference
| Plan | Price | What You Get |
|---|---|---|
| Cursor Hobby | Free | Limited usage |
| Cursor Pro | $20/mo | Composer 2 + Claude + GPT |
| Cursor Pro+ | $60/mo | More usage |
| Cursor Ultra | $200/mo | Unlimited |
| Cursor Teams | $40/user/mo | Team features |
| Claude Pro | $20/mo | Claude Code + API |
| Claude Team | $30/user/mo | Team features |
The Bigger Picture
Cursor’s trajectory in numbers:
| Date | Milestone |
|---|---|
| 2024 | Launched Cursor IDE |
| Nov 2025 | $29.3B valuation, $2.3B funding round |
| Dec 2025 | $1B ARR |
| Feb 2026 | $2B ARR (doubled in 60 days) |
| Mar 2026 | Composer 2 launch, in talks for $50B valuation |
With 1M+ daily active users, 360K+ paying customers, $2B ARR, ~150 employees ($13.3M revenue per person), and now their own frontier-level coding model — Cursor isn’t just a coding tool. It’s becoming one of the most capital-efficient AI companies ever built.
The question is whether they can stay ahead as OpenAI, Anthropic, and Google all build competing coding products. Composer 2 is their strongest argument yet.
FAQ
What is Cursor Composer 2?
Composer 2 is a proprietary AI coding model built by Cursor (Anysphere Inc.), released on March 19, 2026. It’s trained specifically on coding datasets and optimized for Cursor’s agentic IDE workflow, supporting 200K token context windows and multi-step coding tasks.
How much does Composer 2 cost?
Composer 2 Standard costs $0.50 per million input tokens and $2.50 per million output tokens. Composer 2 Fast (the default) costs $1.50 and $7.50 respectively. Both are included with Cursor Pro ($20/month) subscriptions with generous usage pools.
Does Composer 2 beat Claude and GPT?
Composer 2 beats Claude Opus 4.6 on CursorBench, Terminal-Bench 2.0, and SWE-bench Multilingual. GPT-5.4 still leads on Terminal-Bench 2.0 (75.1 vs 61.7). The main advantage is competitive performance at much lower cost.
Can I use Composer 2 outside of Cursor?
No. Composer 2 is only available within the Cursor IDE and the new Glass interface alpha. It’s not offered as a standalone API or foundation model.
Should I switch from Claude Code to Cursor?
Not necessarily. Claude Code excels at complex reasoning and terminal-first workflows. Many developers use both — Cursor for IDE-based work and Claude Code for heavy reasoning tasks. The tools are complementary.
What is Cursor Glass?
Glass is Cursor’s new interface, currently in early alpha. It’s a redesigned experience beyond the VS Code fork, though details are limited. It supports Composer 2 and signals Cursor’s move toward a fully proprietary product stack.
Cursor went from wrapping other people’s models to building its own frontier AI. That’s not just a product update — it’s a company transformation. The AI coding wars just got a lot more interesting.